Use writing structure to detect AI writing, also mansplaining, and more

A new paper on detecting AI writing. The robot summary:

AI-written company blog posts give themselves away by their shape, not their words: they announce a thesis and a roadmap up front, pitch the subject as the modern fix for a legacy way of working, explain in a confident institutional voice and close by restating the thesis. Structure alone catches 97% of them, and having the model reword its own post barely dents that.

Here is what my robot pulled from the paper.

The tells, in order of how strongly they point to AI:

1. The ending restates the thesis or reframes it. “In short, structured onboarding saves time.” This is the single strongest signal in the study.
2. There’s a summary or synthesis section as one of the post’s stages.
3. No way for the reader to take part. No invitation to reply, comment, get in touch, join something or contribute.
4. The thesis and a roadmap come before the first section. “In this post, we’ll cover why onboarding stalls, three fixes that work, and how to measure the difference.”
5. Legacy-versus-modern framing. “Paper checklists were built for a slower era.” The subject is pitched as the modern answer to an old way.
6. Freshness language: “modern,” “next-generation.”
7. A confident institutional voice as the main basis of credibility, rather than first-hand experience, named people or cited evidence.
8. A mechanism-explanation stage, where the post explains how something works in the abstract.

I read a lot of these papers/etc., and using structure to detect AI writing is the new thing in the field. Good instinct.

I am a die-hard conclusion first, Mento memo person. In non-fiction writing, always write the conclusion first. Desert first always, most people won’t make it through the second course.

In doing that, and most parts of the paper, I would avoid “stage direction”: telling the reader what the mechanics are, e.g., “In this paper, we will…” instead take a page from fiction writing: show, don’t tell. Don’t tell me what you’re going to tell me, just tell me.

Also, it’s almost 2026. We are well clear or the woke-revaluation or being talked down to, creeping clear or the Trump and COVID era distrust of experts, and overall more polite about manspla ning and acting like you’re a god handing down wisdom, so avoid “[a] confident institutional voiceas the as the main basis of credibility, rather than first-hand experience, named people or cited evidence.” Ain’t no one got time for that bullshit.

If you want more AI-detection chaff, I would also add, tell your AI to make mistakes and use clunk phrasing sometimes. Perfection is a tell. Like my “In doing that…” paragraph above. What shat-bed that typing it!

Yes, but, the robots adapt

Of course, when I read these kinds of “how to detect AI writing” things, the first thing I do is load up my robot’s style skill and say “adapt to these findings.” It’s a loosing war.

My advice, and SOP, when it comes to AI “slop,” as always, and with all things in life, is: good things are good, and bad things are bad.

I like AI writing, just, as with human writing, not shit-writing.

For example; I read AI reading every day – the summaries of posts and daily briefings of my RSS feeds (they’re like The Economist daily briefings and Harper’s weekly summaries. I mean, after tuning my robot for 3+ years, it writes exactly how I want to read.

Coda: yes, dear reader, I realize I probably broke my own style rules a few times above. (wink)

Original: SlopShape: Identifying AI-Generated Commercial Web Content by Jochen Madler, Sitefire, arXiv 2609.15369v3, September 2026. (PDF)

🔗 AI Blog Posts Have a Detectable Shape: Structure Alone Spots 97% of Them, and Rewording Doesn’t Hide It